
AI-generated summary
Antropic previously operated 'Glasswing' and 'Cyber Verification Program (CVP)' separately, and introduced a new access authority system with the purpose of integrating them and systematizing the security use of AI models.
Some voices of concern... Dimon “Mythos increases global security risk 10 times”
(샌프란시스코=연합뉴스) 권영전 특파원 = 앤트로픽이 인공지능(AI) 모델을 사이버 보안에 활용할 수 있도록 지원하는 '프로젝트 글래스윙'을 확장했다.
앤트로픽은 그간 별도로 운영해온 '글래스윙'과 사이버 검증 프로그램(CVP)을 하나로 통합해 제공한다고 6일(현지시간) 밝혔다.
'글래스윙'은 최상위 AI 모델인 '클로드 미토스'에 대한 접근권이고, CVP는 하위 모델들인 '클로드 오퍼스', '클로드 소넷'의 보안 영역 접근권이었다.
The new access authority system is divided into three levels: 'Defense Access', 'Red Team Access', and 'Special Access'.
방어 접근은 보안센터 운영과 취약점 분석 등 순수 방어 과제를 수행하는 공공기관이나 기업, 대학 등이 대상이다.
레드팀 접근은 시스템에 대한 모의 침투와 공격 테스트 권한 등을 부여받지만, 대규모 혼란을 유발하는 행위는 차단되는 단계다.
최상위 단계인 특수 접근은 전력망, 항공 운항, 통신망, 은행 간 송금망, 정부 행정망 등의 안전장치를 시험할 수 있는 일부 검증기관에만 제공된다.
기존의 글래스윙 참여 기관은 별도 절차 없이 특수 접근 단계로 자동 전환되며, 신규 신청 기관은 미국 정부와 협력해 검증해 특수 접근 자격을 부여한다.
앤트로픽은 지난 4∼7월 글래스윙 프로그램을 통해 12만9천 건 이상의 소프트웨어 취약점을 발견했고, 이 가운데 3만3천 건은 치명적이거나 고위험으로 분류됐다고 소개했다.
However, there are concerns raised about the improvement in cyber security capabilities of such AI models.
제이미 다이먼 JP모건체이스 최고경영자(CEO)는 이날 블룸버그TV와 인터뷰에서 "앤트로픽의 미토스 모델이 전 세계 사이버 보안 위험을 10배 높였다"고 지적했다.
AI outlook — possibilities, not facts
The U.S. government will expand cooperation with Antropic to gradually increase the number of verification agencies eligible for special access.
Likely · Within months
As demand for AI-based security verification increases in key infrastructure fields such as finance and energy, the related market size is expected to expand.
Possible · Within months
![[AI Prism] Domestic AI companies “toward an open ecosystem”... Lisa Su “Let’s go together” (Comprehensive 2nd edition)](/api/img?u=https%3A%2F%2Fimg.yna.co.kr%2Fphoto%2Fyna%2FYH%2F2026%2F10%2F07%2FPYH2026100704790001300_P2.jpg&w=320&q=72&f=webp)
AMD Chairman Lisa Su announced that she will cooperate with 13 domestic AI semiconductor companies to build a heterogeneous AI infrastructure that combines CPU/GPU and domestic AI semiconductors, develop it into a global reference model, and support the establishment of an AI research base in Korea and participation in the ROCm ecosystem.

Bloomberg reported that Chinese AI company DeepSeek is expected to secure close to 100 billion yuan (about 20 trillion won) in funding in its ongoing investment round. CATL and Tencent participated as major investors, and Deepseek's corporate value is expected to reach at least 500 billion yuan (about 100 trillion won).
![[AI Pick] Lisa Su “Betting on Korea’s AI ecosystem”… Recruiting hundreds of researchers](/api/img?u=https%3A%2F%2Fimg.yna.co.kr%2Fetc%2Finner%2FKR%2F2026%2F10%2F07%2FAKR20261007049900017_01_i_P2.jpg&w=320&q=72&f=webp)
AMD CEO Lisa Su visited Korea and announced plans to establish an AI research base (CoE) in Korea, promising to hire hundreds of researchers and expand infrastructure and industry-academic cooperation with domestic companies and universities. Cooperation discussions with 14 companies, including Rebellion, Furiosa AI, and Mango Boost, are also scheduled.

Finland's Licensing Supervisory Authority ordered Google to stop work on data center site development in the Muhos and Kayani regions until an environmental impact assessment is completed. Google planned to invest at least 13 billion euros in Finland over the next two years, but there are concerns that this measure will delay the project.

Apple is collaborating with LG Electronics to jointly develop smart home devices, which will be manufactured and sold under the LG brand, while Apple will participate in design and function development. The product will be linked to Apple's new smart home hub, which is interpreted as a strategy to challenge Amazon's Ring and Google's Nest.

A Wall Street Journal report revealed that ARTEX, an open source cybersecurity tool developed in China, was abused to hack a Korean financial company. Hackers used this tool to carry out security attacks using multiple AI models, and the attacks were carried out through more than 10 countries, including the United States, Japan, and Germany. The developer specified user guidelines to prevent abuse, but it was pointed out that its effectiveness was limited.